Heterogeneous data fusion considering spatial correlations using graph convolutional networks and its application in air quality prediction

Models that predict the future state of certain observations are commonly developed by utilizing heterogeneous data. Most traditional prediction models tend to ignore inconsistencies and imperfections in heterogeneous data, and they are also limited in their ability to consider spatial correlations among monitoring points, as well as predict for the entire study area simultaneously. As a solution, this paper proposes a deep learning method to fuse heterogeneous data collected from multiple monitoring points by exploiting graph convolutional networks (GCNs), thus perform prediction for certain observations. The effectiveness of this approach was evaluated by applying it to a prediction scenario for air quality. As the fundamental idea behind the proposed method, (1) the collected heterogeneous data is fused according to the coordinates of the monitoring points considering its spatial correlations, and (2) the prediction considers global information as opposed to local information. According to an air quality prediction scenario, (1) the fused data obtained by the RBF-based fusion method are reasonable and reliable; (2) the fusion significantly increases predictive model performance; and (3) the STGCN model enhanced by the fusion achieves the highest performance. This approach can be similarly applied in scenarios involving continuous heterogeneous data collected from scattered multiple monitoring points in the study area.